Maximal Arc-Length Matching at Multiple Thresholds (MALT) for Variational Shape Recognition

Widely varying shapes are diÆcult to recognize especially when present as part of an occluded shape. Handling such situations requires the consolidation of data at di erent levels of con dence, which is the philosophy motivating the Maximal Arc-Length matching at multi-Thresholds algorithm (MALT). The algorithm matches segments from two contours and rates a pair of matches more highly if the arc-length shift in the two matches is equal. Unlike traditional matching strategies, dependency on threshold selection is avoided by considering multiple thresholds and ranking the matches. All the models are tested in the polygonal domain, in which variational shape classes at di erent levels of deformation have been generated using vertex and edge deformation from a nominal model. Although in its current form the algorithm is restricted to polygonal contours, in this domain it exhibits very robust recognition even in the presence high deformations and occlusion.